I am a Masters student in Computer Science at Stanford University with Artificial Intelligence and Systems specialization and a graduate of Electrical Engineering and Computer Sciences at UC Berkeley.
Previously, I spent over two years as a Machine Learning Researcher at the PALLAS Group at Berkeley Artificial Intelligence Research Lab (BAIR), advised by Professor Kurt Keutzer.
Currently, I am working as a Machine Learning Engineer at IntuigenceAI, where I finetune synthetic Engineering models and design and deploy multi-agent LLM systems for industrial applications.
My research focuses on efficient deep learning, particularly for Large Language Models. I am interested in KV cache quantization, speculative decoding, and building scalable AI agent systems. I have published at venues including ACL, NeurIPS, and COLM. I like working on both the algorithms and the systems they run on.
Stanford, CA
--lean mode for spawning peers. New landing page at hivam.org/cxpeer.
A ten-part series that takes modern language models apart, the transformer, BERT, GPT-2, T5, and LLaMA, down to their weight matrices, tensor shapes, and FLOP counts. Every claim is one you can check with arithmetic. Written from my UC Berkeley CS199 study and hand-illustrated throughout.
Read the series →
Codex CLI sessions as Claude Code peers. A Claude session sends a task with SendMessage, Codex works in its own context, and the answer comes back when the turn ends. Built on Claude Code's undocumented peer registry and Codex's hooks: three files, one Unix socket, no daemon. Includes a context meter that reads each session's live prompt size from the transcripts both tools write.
Landing page and demoA staff-engineer AI judge runs design interviews and production incidents against a hire bar, a persistent learner model tracks every concept, and a fine-tuned router picks the tutor's context under a token budget. The router was trained as LoRA on four Qwen2.5 sizes on a 32-GPU H100 cluster; the 7B one matches the frontier oracle's selection at 0.81 Jaccard.
An OpenAI-compatible serving engine in the spirit of mini-sglang: a bare-bones Qwen3 in pure PyTorch with GQA, RoPE, and SwiGLU, prefill and batched decode, top-k and top-p sampling, an SSE-streaming FastAPI server, and a scheduler written to be the optimization target.
An implementation of ReasoningBank: a ReAct agent with browser tools that extracts reasoning strategies from its own trajectories, stores them in an embedding memory bank, and retrieves them for new tasks. Includes an LLM judge for trajectories and an evaluation harness for WebArena, SWE-bench, and Mind2Web.
An FPGA datapath with runtime-selectable microscaling formats for energy-efficient RLHF training. Profiling all 169 quantizable layers of Qwen2.5-0.5B showed 93.5% tolerate MXFP4; the rest fall back to MXFP8.
Multi-agent LLM systems on domain-specific fine-tuned models, context-aware retrieval pipelines, and inference deployment on Azure AI Studio.
Python solvers for nonlinear differential equations that made analytics 75% faster; simulated home appliance power draw from Span Panel data for the next product iteration.
Efficient LLM inference: co-author of Squeezed Attention (ACL 2025), KVQuant (NeurIPS 2024), and SPEED (NeurIPS 2023 workshop).
Fine-tuned and prompt-tuned language models for scientific-article chatbots as an individual course of study.
Built and analyzed a mantle model of Venus on supercomputers in Python and Fortran; co-authored a paper supporting NASA's VERITAS mission.
For TensorZipper, a model compression algorithm.
Two-time recipient.